Skip to the content
Krizaka Blog
The Krizaka story

AI

Is that AI model really open source? What the licence actually says

Open source, open weights, source-available: what these words promise, what a licence really controls, and what I read while choosing the image, video and voice models behind our products.

A green desk lamp and a magnifying glass over a plan with small red circles drawn on it

On 15 October I checked that none of the default models in our AI engine carried a non-commercial licence, since we sell to companies. The default English voice was called Ryan. It was a Piper voice, free, published "as open source". When I opened its card, I read: CC BY-NC-SA 4.0. The "NC" means non-commercial.

The program is free, the model weights can be downloaded, everyone says "open source". But the voice was trained on a dataset that forbids commercial use, and that restriction follows the voice. Ryan stayed available, never as a default, and a public-domain voice took its place.

Since then I no longer choose a model on the word "open". I read the licence. Here is what I learned, with examples I opened myself at the source. And right away: I am an engineer, not a lawyer. This post is an engineer's reading, not legal advice. For a decision that binds your company, have someone whose job it is read the licence.

Three words people mix up

Open source has a precise meaning. The Open Source Initiative (OSI) defines what a licence must allow: use, study, modify and share, for any purpose. Its definition even forbids discriminating against a field of endeavour. A licence that says "not for this industry" is not open source, whatever the label.

Open weights means something more modest: the file of the model's parameters can be downloaded. You can run it at home. But the training data and the training code are not necessarily there, and the licence can set conditions. The OSI wrote a whole piece about it, "Open Weights: not quite what you've been told".

Source-available describes what you can read without having the right to do everything with it. Handy for understanding, less so for building a product.

In October 2024 the OSI published its Open Source AI Definition (OSAID 1.0). It takes the four freedoms of free software and asks for three things from an AI system: the parameters, the code to train and run it, and enough detail about the data that a skilled person could rebuild an equivalent system. By that rule, many models called "open" are open-weights models. They are useful, often generous, but it is not the same promise.

What a licence can control

We first think of "commercial or not". It is the best known, but not the only one.

A usage threshold: the licence of Llama 3.1 asks you to request a licence from Meta if your products pass 700 million monthly active users. A revenue threshold: LTX-Video's asks for a paid licence from 10 million dollars in annual revenue.

A field of use: the Llama licence incorporates an acceptable use policy, a list of what is forbidden, which you must follow. A geography: as we will see, a licence can exclude whole countries. The outputs: some licences say what you may do with what the model produces, for example forbidding you to use it to train another model. Attribution: Llama asks you to display "Built with Llama". Redistribution: if you fine-tune a model, what must you pass on, and under which terms? And the acceptance gate on Hugging Face, where you must click to get the files. Clicking is signing.

What I read while choosing our models

Each time I opened the licence file or the model card, not an article about it.

Video. The licence of HunyuanVideo opens with a sentence in capitals: it does not apply in the European Union, the United Kingdom and South Korea. Further down, it also forbids using the model's outputs outside the covered territory, and using them to improve another model. Our site is public: anyone in Europe can open it. It was dropped in our video bake-off, and the licence is the reason we wrote down.

Wan 2.2 A14B is under Apache 2.0, an OSI-approved licence, with no threshold and no territory. It makes our clips. LTX-Video, in its 0.9.8 version, has a revenue threshold of 10 million dollars. We are very far below it, so we can use it for previews. But the licence also asks you to pass its use restrictions on to your own users. I did not keep LTX-2: the disk space, and the same revenue condition.

Images. With FLUX.1 [dev], the licence is non-commercial: you cannot run the model for a commercial activity, even though the images it produces can be used for anything, and content filters are required. FLUX.1 [schnell], from the same publisher, is under Apache 2.0. Same family, two regimes. It continues with the next generation: FLUX.2 [klein] 4B is under Apache 2.0, but FLUX.2 [klein] 9B is under a non-commercial licence. We use the 4B. To raise resolution, SeedVR2 is under Apache 2.0.

Voices. This is where I learned the most. XTTS-v2 is under Coqui's CPML licence, which, in its own text, "allows only non-commercial use of a machine learning model and its outputs". Its code is under MPL 2.0, though; the model's licence is another one. And the page the model card cites to read it, coqui.ai/cpml, answers 404 today: the text now only exists in the repository's file. Keep a copy of the licence at the moment you choose. Kokoro is under Apache 2.0, and we use it.

Piper shows that a model's licence can come from its data. The voices of the piper-voices repository each have their own card: Ryan under CC BY-NC-SA 4.0, Siwis (French) under CC BY 4.0, LJSpeech (English) in the public domain, Tom (French) under AGPL-3.0. Since then, the model catalogue has a commercial_use column (yes, with conditions, no), filled from these cards.

Audio. MMAudio, which invents the sound of a video, is published under CC BY-NC 4.0: non-commercial, so not for a product. The music in our clips comes from ACE-Step, under MIT. Stable Audio Open is under Stability AI's community licence: you have to register, and it is free only below one million dollars of annual revenue.

ModelLicenceCommercial useGood to knowHere
HunyuanVideoTencent Hunyuan CommunityOutside the EU, UK, South KoreaOutputs included in the restrictionDropped
Wan 2.2 A14BApache 2.0YesNoneUsed
LTX-Video 0.9.8LTX Open WeightsUnder $10 M revenueRestrictions to pass onUsed for previews
FLUX.1 [dev]FLUX.1 [dev] non-commercialNot for the modelOutputs usable, filters requiredNo
FLUX.1 [schnell]Apache 2.0YesNonePossible
FLUX.2 [klein] 4BApache 2.0YesNoneUsed
FLUX.2 [klein] 9BFLUX non-commercialNoFilters requiredNo
SeedVR2Apache 2.0YesNoneUsed
XTTS-v2CPML (Coqui)NoText offlineNo
KokoroApache 2.0YesNoneUsed
Piper, voice by voiceCC BY, CC BY-NC-SA, public, AGPLDepends on the voiceData mattersSiwis, LJSpeech
MMAudioCC BY-NC 4.0NoNoneNo
Stable Audio OpenStability AI CommunityUnder $1 M revenueRegistration requiredNo

The weights are only one piece

A model is not just a file. When you run it, there is the inference code, with its own licence: Piper's engine, for example, was under MIT, its original repository is archived, and its successor is under GPL-3.0. There are the libraries all of it depends on, each with its own. There is the training data, whose licence can travel up to you, as with Ryan. And there is where the model runs: on your machine, you are bound by the model's licence; through a provider's API, you are also bound by its terms of use, which can be stricter. I wrote about it in the post on uncensored models.

My list before choosing

  1. Open the model's licence file, not its summary. Copy it with the date.
  2. Search for the words: commercial, revenue, users, territory, outputs, forbidden.
  3. Read the acceptable use policy, if there is one, and ask whether your customers will respect it.
  4. Check the territory: where are your users?
  5. Look at the licence of the data or of the exact variant (the voice, the size, the version).
  6. Read the licence of the inference code and of the dependencies.
  7. Write down the revenue threshold, and the day you will pass it.
  8. If you fine-tune the model, read what you must pass on to your own users.
  9. If you go through an API, read its provider's terms too.
  10. Write the result next to the model's name. I do it in a column, for each one. It ties in with what I said in the post on providers: the licence comes before the price.

The licences, one by one

ModelUse hereLicence of the weights
Wan 2.2 I2V A14BCharacter takesApache 2.0
LTX-Video 0.9.8LandscapesLTX Open Weights
FLUX.2 [klein] 4BKeyframesApache 2.0
SeedVR24K upscaleApache 2.0
ACE-StepMusicMIT
Zonos v0.1NarratorApache 2.0
KokoroSynthetic reference voiceApache 2.0
WhisperChecking the spoken wordsMIT
ChatterboxTried, not keptMIT
F5-TTSRefusedNon-commercial (CC BY-NC 4.0)
XTTS-v2RefusedNon-commercial (CPML)
MMAudioRefusedNon-commercial (CC BY-NC)
Stable Audio OpenRefusedCommunity licence, registration required
HunyuanVideo (foley)Set asideExcludes the EU, the UK and South Korea

What to take away

"Open source" is often a label, and the licence is the contract. Two models of the same family, in the same week, can have very different rights. Read the file, not the article, write down what you read, and do it again when the model changes version.

For what it costs to run these models, see the post on model families and the GPU and the one on the real cost of an image.

Keep it, pass it on

Related posts

AI · 12 min read

Models, their families, and why you need a GPU

Dense or MoE LLMs, embeddings, vision, image and video diffusion, audio, upscalers: the model families it took me months to untangle, and why each one has to fit in a GPU's memory. With our measured times.

AI · 8 min read

Understanding AI before you start: the words, without the jargon

Token, context window, quantization, RAG, MCP, agent, diffusion. The words I use without explaining them, taken one by one with a simple picture and what really runs on our side.

AI · 7 min read

The anatomy of an AI prompt: what each added line really changes

A six-word prompt, then three layers of detail, on my Mac with the same seed. What gets better, what gets worse, what a negative prompt can do, and what I didn't prove.